Fangmin Lv created ZOOKEEPER-3144:
-------------------------------------
Summary: Potential ephemeral nodes inconsistent due to global
session inconsistent after with fuzzy snapshot
Key: ZOOKEEPER-3144
URL: https://issues.apache.org/jira/browse/ZOOKEEPER-3144
Project: ZooKeeper
Issue Type: Bug
Components: server
Affects Versions: 3.4.13, 3.5.4, 3.6.0
Reporter: Fangmin Lv
Assignee: Fangmin Lv
Fix For: 3.6.0
Found this issue recently when checking another prod issue, the problem is that
the current code will update lastProcessedZxid before it's actually making
change for the global sessions in the DataTree.
In case there is a snapshot taking in progress, and there is a small time stall
between set lastProcessedZxid and update the session in DataTree due to reasons
like thread context switch or GC, etc, then it's possible the lastProcessedZxid
is actually set to the future which doesn't include the global session change
(add or remove).
When reload this snapshot and it's txns, it will replay txns from
lastProcessedZxid + 1, so it won't create the global session anymore, which
could cause data inconsistent.
When global sessions are inconsistent, it might have ephemeral inconsistent as
well, since the leader will delete all the ephemerals locally if there is no
global sessions associated with it, and if someone have snapshot sync with it
then that server will not have that ephemeral as well, but others will. It will
also have global session renew issue for that problematic session.
The same issue exist for the closeSession txn, we need to move these global
session update logic before processTxn, so the lastProcessedZxid will not miss
the global session here.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
